Switch StdRng and thread_rng to HC-128

[pitdicker]

As discussed in dhardy#53.

I also changed THREAD_RNG_RESEED_THRESHOLD to 32 MiB, which seems a more reasonable number to me.

[dhardy]

A couple of little tweaks needed, but I approve, despite the apparent performance issue on ARMv7. Feedback in the linked thread was positive

[dhardy]

Doesn't this leave IsaacWordRng unused? We might as well remove it. Ah, you didn't update the Seed type...

[dhardy]

I guess this means it can implement CryptoRng now.

[pitdicker]

I was too busy with the docs that I forget to check the rebase 😕. I will update the tests in a moment, but am getting some strange failures...

A couple of little tweaks needed, but I approve, despite the apparent performance issue on ARMv7.

I would like to try a couple of benchmarks on ARM, but am not sure how to set it up.

[pitdicker]

Wow, testing on my phone is super easy with cargo dinghy test!
rand still passes all tests on ARM 😄. But cargo dinghy bench only runs the benchmarks as test, without performance numbers. So that will need some more experimenting.

[dhardy]

Use of nautical terms reminds me that my phone is a Jolla, with native shell and GCC support... Rustup is also happy to install. Slow but managed to get some results at least:

test gen_bytes_chacha12    ... bench:  10,633,598 ns/iter (+/- 61,043) = 96 MB/s
test gen_bytes_chacha20    ... bench:  16,384,802 ns/iter (+/- 48,325) = 62 MB/s
test gen_bytes_chacha8     ... bench:   7,751,370 ns/iter (+/- 61,750) = 132 MB/s
test gen_bytes_hc128       ... bench:   4,181,004 ns/iter (+/- 21,645) = 244 MB/s
test gen_bytes_isaac       ... bench:   2,996,947 ns/iter (+/- 18,311) = 341 MB/s
test gen_bytes_isaac64     ... bench:   2,478,126 ns/iter (+/- 18,311) = 413 MB/s
test gen_bytes_xorshift    ... bench:   2,661,644 ns/iter (+/- 297,298) = 384 MB/s

Funny when Isaac64 beats Xorshift.

[pitdicker]

Funny when Isaac64 beats Xorshift.

Sounds like there is some work left to do on ARM...

Travis seems to have given up, but because this passes on all other platforms I can't imagine it to fail. Ready to merge?

[dhardy]

I need to investigate those benchmarks further; some of them seemed to hang.

Oh you still didn't implement CryptoRng? That was added as part of #273 I think.

[dhardy]

Oh, Xorshift wins on u32 and u64. Tests sometimes hang but when they do give a result, the results do appear to be repeatable. Isaac still beats HC128 everywhere in performance, but not by too much.

test gen_bytes_chacha12    ... bench:  10,666,902 ns/iter (+/- 71,778,443) = 95 MB/s
test gen_bytes_chacha20    ... bench:  16,412,455 ns/iter (+/- 38,287,657) = 62 MB/s
test gen_bytes_chacha8     ... bench:   7,772,730 ns/iter (+/- 48,848) = 131 MB/s
test gen_bytes_hc128       ... bench:   4,182,483 ns/iter (+/- 13,923) = 244 MB/s
test gen_bytes_isaac       ... bench:   3,004,071 ns/iter (+/- 54,952) = 340 MB/s
test gen_bytes_isaac64     ... bench:   2,460,653 ns/iter (+/- 55,563) = 416 MB/s
test gen_bytes_xorshift    ... bench:   2,652,987 ns/iter (+/- 68,843) = 385 MB/s

test gen_u32_chacha12      ... bench:      47,129 ns/iter (+/- 381) = 84 MB/s
test gen_u32_chacha20      ... bench:      69,453 ns/iter (+/- 763) = 57 MB/s
test gen_u32_chacha8       ... bench:      35,758 ns/iter (+/- 172,155) = 111 MB/s
test gen_u32_hc128         ... bench:      20,893 ns/iter (+/- 60) = 191 MB/s
test gen_u32_isaac         ... bench:      18,317 ns/iter (+/- 1,097) = 218 MB/s
test gen_u32_isaac64       ... bench:      16,791 ns/iter (+/- 763) = 238 MB/s
test gen_u32_xorshift      ... bench:       9,254 ns/iter (+/- 95) = 432 MB/s

test gen_u64_chacha12      ... bench:      89,727 ns/iter (+/- 1,253) = 89 MB/s
test gen_u64_chacha20      ... bench:     135,030 ns/iter (+/- 1,961) = 59 MB/s
test gen_u64_chacha8       ... bench:      67,148 ns/iter (+/- 1,145) = 119 MB/s
test gen_u64_hc128         ... bench:      37,398 ns/iter (+/- 381) = 213 MB/s
test gen_u64_isaac         ... bench:      28,811 ns/iter (+/- 763) = 277 MB/s
test gen_u64_isaac64       ... bench:      23,850 ns/iter (+/- 1,526) = 335 MB/s
test gen_u64_xorshift      ... bench:      10,685 ns/iter (+/- 190) = 748 MB/s

[pitdicker]

Those benchmarks are very strange, gen_u64_xorshift should have roughly the same Mb/s as gen_u32_xorshift. But thanks for testing, and for implementing CryptoRng.